Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷

Machine Learning Master's student at @UTokyo_News_en. Staying at @NTUSg from Jul-Dec '26. Intern at EQUES, @Matsuo_Lab. Prev: @MatsuoInstitute, IS23er.

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

Our paper:
https://t.co/k9Yr4teESz

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

Come to see our poster today!

🏟️Hall A, #3106
🕐4:45pm-5:30pm (but if you want to chat at a different time, drop me a message anytime via DM or ICML App!)

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

We present our paper "Mitigating Reward Hacking via Adversarial Robustness" at EIML@ICML2026!

We conjecture that reward hacking is often caused by flipped advantage-sign estimations, and propose SignCert-PO, a new algorithm built on the theory of randomized smoothing! 🧵

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

@pcastr Sooo wet but I still had a lot of fun, thank you!

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

Update: I’m honored to have been selected as a Toyota Riken Overseas Predoctoral Fellow (OSPDF), which provides substantial support including tuition and living expenses.

I’m actively looking for Fall 2027 PhD positions in the following areas.

I’d be very happy to chat at ICML!

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

Thank you Johannes, it has been a real pleasure working with you!

I'm currently looking for PhD positions for Fall 2027. I'm with an ML background, interested in AI alignment, multi-agent RL, strategic decision making and more.

I'll be at ICML and would be very happy to chat!

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

@pcastr @alexUnder_sky @changrybirds I'll join you tomorrow!

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

@pcastr Can't make it tomorrow, but I'l love to join if there's another on or after Wednesday :)

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

@PandaAshwinee DM'd! :)

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

@ishapuri101 DM'd! :)

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

@grx_xce Love to chat about real-world eval!

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

Thank you Johannes, it has been a real pleasure working with you!

I'm currently looking for PhD positions for Fall 2027. I'm with an ML background, interested in AI alignment, multi-agent RL, strategic decision making and more.

I'll be at ICML and would be very happy to chat!

[Johannes Ackermann @johannesack]

The lead author, @shinnosukeono, is also currently looking for PhD positions!

He is one of the most talented master's students I have had the pleasure of working with, I highly recommend him!

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

Come chat on Friday, July 10 at EIML@ICML2026, Seoul!🇰🇷

Joint work with @johannesack, @nissymori1, @tksii and Masashi Sugiyama.

https://t.co/k9Yr4teESz

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

We present our paper "Mitigating Reward Hacking via Adversarial Robustness" at EIML@ICML2026!

We conjecture that reward hacking is often caused by flipped advantage-sign estimations, and propose SignCert-PO, a new algorithm built on the theory of randomized smoothing! 🧵

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

Across model sizes and benchmarks, SignCert-PO performs well and prevents reward hacking. It holds back the policy from going unreliable regions, and this is done dynamically during optimization!

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

By the way any #runconference sessions during #ICML2026?

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

We show SignCert-PO holds back the policy from going "uncertain" regions and prevents reward hacking!

[Shinnosuke ONO / 小野新之介 ✈️ ICML2026 🇰🇷 @shinnosukeono]

Then, we propose SignCert-PO, a new algorithm to handle this uncertainty. Using the theory of randomized smoothing, we show our algorithm results in a surprisingly simple form!